Polyethylene (PE) pipelines with electrofusion (EF) joining is an essential method of transportation of gas energy. EF joints areweak points for leakage and therefore, Nondestructive testing (NDT) methods including ultrasonic array technology are necessary.This paper presents a practical NDT method of fusion joints of polyethylene piping using intelligent ultrasonic image processingtechniques. In the proposed method, to detect the defects of electrofusion joints, the NDT is applied based on an ANN-Waveletmethod as a digital image processing technique. The proposed approach includes four steps. First an ultrasonic-phased arraytechnique is used to provide real time images of high resolution. In the second step, the images are preprocessed by digital imageprocessing techniques for noise reduction and detection of ROI (Region of Interest). Furthermore, to make more improvement onthe images, mathematical morphology techniques such as dilation and erosion are applied. In the 3rd step, a wavelet transform isused to develop a feature vector containing 3-dimensional information on various types of defects. In the final step, all the featurevectors are classified through a backpropagation-based ANN algorithm. The obtained results show that the proposed algorithmsare highly reliable and also precise for NDT monitoring.
1. Introduction
The ultrasonic technique as a nondestructive testing (NDT)method has been widely used over decades to evaluate thequality of materials and equipments without causing dam-age in a large range of industries. In the evaluation of pressurevessels and piping, not only is UT utilized in manufacturingquality controlling, but also has been used in service moni-toring and residual life prediction, such as the inspection ofwelded joints, monitoring of crack propagation, and evalu-ation of materials property deterioration. In the specific caseof welded materials, the research for the development of anacceptable system for analyzing the extracted images fromthe welded joints has grown considerably in the last years [1–4].
One of its applications is in the gas pipelines where theusage of natural gas in residential, commercial, and industrialfacilities is increased day by day. In this way, Polyethylenepipes rapidly substituted the metal pipes, because the poly-ethylene pipes have a high-corrosion resistance, easy to form,lighter, and cheaper than metal ones. In fact, the main reason
for using PE pipes in gas distribution is that its materialhas a high-chemical resistance against corrosive materials intransported gas. In addition, PE pipes are easy to carry, liedown, and make connections. Because of these benefits, gasdistribution companies and water and sewage organizationswould change their existing systems and use the PE pipes.The demand of polyethylene (PE) pipeline is increased forgas energy transportation and electrofusion (EF) joining isan essential method to build PE pipeline [5].
It is important to note that, usually the EF joints are con-sidered as weak points for leakage, and kind of nondestruc-tive test is necessary. One of the main factors disturbing thereliability and accuracy of the test is the encountered noiseduring inspection. The most commonly used ultrasonicdetector is the A-scan detector. This kind of traditional UThas several disadvantages such as the need for a skilled andexperienced technician to judge the defect, also the lack ofpermanent record, which is extremely important in the con-dition monitoring and in-service inspection [5].
These problems may be easily solved by the introductionof a digital ultrasonic system, which combines the computer
2 Journal of Control Science and Engineering
Terminalconnector
Main pipe
EF coupler
Main pipe
Figure 1: (A) Electrofusion (EF) joining of two polyethylene (PE)pipes using EF coupler. (B) Cross-section of EF coupler includingheating wires.
and digital signal processing (DSP) technology into UTinstrument [6–9]. Electrical, pulse, ringing, and structurenoises are the most commonly encountered noises, whichreduce the quality of extracted images for NDT evaluations.But in order to improve the process of image processing,wavelet transforms are used significantly [10–12]. The wave-let transform, multiresolution analysis, and other space fre-quency or space scale approaches are now considered asstandard tools by the researchers in signal processing, andmany applications have been proposed. The theory of wave-let analysis has been well studied in [13], where images arerepresented by wavelet reconstructing toolbox.
Bhuiyan et al. have successfully used wavelet transformsto significantly improve identification and classification rateson fingerprints. They showed that wavelet contains featuresthat are more pronounced for higher accuracy in recognizingfingerprints [10].
Also in [11, 12] the output of wavelet transformation toobtain some features is used as an input to the artificial neu-ral network (ANN) classifier for pattern recognition.
The aim of this paper is to present a new combined intel-ligent algorithm considering digital image processing, neuralnetworks, and mathematical morphology techniques forimproving the quality of extracted images from NDT evalu-ations. Features are extracted from Haar wavelet decomposi-tion of the JPG images. The simulations are done usingMATLAB platform. Obtained results show that the proposedalgorithms to ultrasonic signals are highly reliable and pre-cise for quality of NDT testing and monitoring.
2. Electrofusion Joining of PE Pipe
EF joining is one of the widespread PE pipe weld methods.An estimated annual use of EF joining was over 15 millionin 1993 [1]. This technique makes possible joining of pre-assembled pipes to be carried out with minimum equipment.Electrofusion method is a system that welds pipes togetherthrough fittings whose internal surfaces are covered withspecial resistance wires (as shown in Figure 1). Welding isperformed through melting plastic material with heatingcoils that reach a high temperature as a result of the stressapplied to the sockets on fittings by an electrofusion mach-ine. The electrofusion welding process can be described
Manufactures labelling
Lack of penetrationLack of fusion
Void
Figure 2: General structure and location of joint flaws.
in three stages: initial heating and fitting expansion, heatsoaking to create the joint, and finally joint cooling. Theapplication of the electrofusion process is shown in Figure 1.
Usually, wound-heating wires exist between the couplerand main pipes. The distance between adjacent wires is usu-ally very close, for example, 1∼3 mm. The fundamental ideasof joining process are to heat the wires and to melt the insideand outside surface of coupler and main pipes respectively,and then to consolidate fusion area. Welding faults (badpreparation of the pipe, poor cleaning, pipe surface being notscraped, pipe and fitting being badly clamped, or not respect-ing of fusion time) can generate defects. Figure 2 showsa cross-section view of the coupling. The position of theheating wires and possible flaw locations are shown.
3. Proposed Algorithm
In order to improve the raw UT images for observation andaccurate analysis, a combined algorithm is proposed con-sidering various image preprocessing, mathematical mor-phology (such as dilation and erosion operations), waveletfeature extractor, and artificial neural network. Flowchart ofproposed algorithm is shown in Figure 3.
The starting point of this algorithm is collecting andinserting the raw UT images. In this way, all electrofusion(EF) joints of polyethylene pipes must be tested by ultrasonicmethod. Then extracted images will be prepared and savedas JPEG pictures for computer evaluations.
Once the template and test images are resized, thresholdvalues for the grayscale images are determined to convertthe images into binary ones. Grayscale images with levelsbetween 0 and 255 are converted into binary images. It isassumed that the initial threshold is equal to 0.55. This valuewill be updated during this algorithm if it cannot find anappropriate result.
Then binary images will be fed into the image processingstep including preprocessing sections such as noise reductionand segmentation based on the image processing toolbox
Journal of Control Science and Engineering 3
Make binary with 0.55 threshold
Preprocessing
Input raw UT
Haar wavelet transform
Feature extraction
ANN
Binary with fewer thresholds
(Threshold <0.4)
Yes
No
Boundary calculation
Indicate circular index of objects
Find objects with
Resizing and upsampling
circular index >0.83
Interested image (ROI)
The number of objects = 0
Figure 3: Schematic diagram of proposed algorithm.
of MATLAB software [14]. Preprocessing involves applyingradiometric and geometric corrections to a raw image data.Several levels of corrections have been defined. For thepreprocessing step, all noises and unwanted objects will beremoved from the document’s image. This leads to an easierand more effective process. Afterward, more improvementwill be made bymorphological operations.
Mathematical morphology [15] is a mathematical toolfor analyzing image on morphological basis. Its basic ideais to use structuring elements of certain morphology tomeasure and extract corresponding morphology in the imagefor making analysis and recognition. The application ofmathematical morphology helps to simplify image data, sothe basic morphologies of image can be maintained andunrelated structure can be removed. Mathematical morphol-ogy consists of four basic steps including morphology corro-sion, morphology expansion, opening operation, and closingoperation. In this paper, we focus on the dilation and erosionoperations of the mathematical morphology. Given f (x, y) isa grey-scale image on domain Z. g(x, y) is a structuring ele-ment on domain Z hence, the formula for grey-scale expan-sion operation of image f (x, y) based on g(x, y) is as follows:
(fΘg
)(x, y
) = min(f(x + i, y + i
)− g(i, j))
,(i, j) ∈ Dg ,
(x, y
) ∈ Df .(1)
Accordingly, the erosion of f (x, y) by g(x, y), denoted byfΘg, is defined as
(f ⊕ g
)(x, y
) = max(f(x + i, y + i
)+ g(i, j))
,(i, j) ∈ Dg ,
(x, y
) ∈ Df .(2)
Many other morphological operations are based on thesetwo basic operations [15].
The operations of erosion, dilation, opening, closing,and others can extract many types of information about abinary image. Morphological reconstruction can be appliedfor restoring the lost image information and also for seg-menting object image as shown simulations [16]. Then theprepared ROI section from the images will be used by wave-let transform. Finally, all features like mean and standarddeviation of data will be trained by a multilayer neuralnetwork. Details regarding this algorithm and its simulationare presented in the next section.
4. Simulation Results
The experimental results and performance evaluation ofthe proposed method are described in this section. Realtime ultrasonic array technique was applied to obtain theultrasonic images of the cross-section of electrofusion joints.
4 Journal of Control Science and Engineering
Ultrasonic array transducers in this simulation have 96 arrayelements and the center frequencies of the transducers were7.5 MHz for the high-resolution application. To monitor thereal time image, a PC monitor is connected to the system.Also 3D-FFT (Haar wavelet) was used in feature extractionprocesses of ROI images and simulations were performed on2 GHz PC by using MATLAB.
4.1. Image Resizing. Tests showed that the best performancewill be achieved by increasing the dimension of inputpictures to twice their initial size. Because the raw image aresmall and by increasing the dimension more details such aspixels and the edges of regions will appear significantly. Theresults of this part are presented in Figure 4.
4.2. Noise Reduction and Morphological Application. Noisereduction is the process of removing noise, holes, andunwanted details from images. Noises in binary imagesmainly consist of isolated pixels of opposite value in imageobjects or background (also called salt-and-pepper noise);small holes on objects and small spots on the background;line merging and splitting. In the simulated case of this paper,a combined binary filter is used for erosion or dilation ofobjects, removal of noise in an image, detect edges, andsmoothing the image.
This filter consists of two steps. In first step, a separateprogram is applied to indicate and remove all holes withareas less than 30 (This threshold is determined by text).Then, in the second step, morphological method is appliedfor more improvement. The result of noise reduction basedon first step of binary filter is presented in Figure 5. It is clearthat unwanted areas are removed from the images.
But in case of morphological application, designing astructural element is necessary. Mathematical morphologyregards an image as a set and uses another smaller set whichis called as structuring element to probe the image.
This apparent geometric description of set theory makesmathematical morphology more suitable for visual infor-mation processing. Mathematical morphology is originallyproposed for binary images, and its basic theory is developedin this application. In this case, a 5 × 5 structuring element(Figure 6) is designed for morphological application. So insome cases, the results of this application are shown inFigure 7.
Based on the proposed algorithm in Figure 3, the impor-tant step of this algorithm is to extract a high-quality ROI forfurther examination by wavelet transform. So it is necessaryto trace and indicate the exterior boundaries of objects, aswell as boundaries of holes inside the objects, in the binaryimage. So nonzero pixels will belong to an object and 0 pixelsconstitute the background. These boundaries for each regionin binary images are specified using MATLAB functions. Inthe next step, circular index of each obtained object fromprevious step should be indicated. Circular index is definedby
Circular Index = 4× π × S
P2, (3)
where S and P are the area and perimeters for each object,respectively.
To imply this matter, an acceptable value for circularindex (circular threshold) should be found for a sufficientcomparison with current situation of each object.
Tests and simulations on database showed that the bestvalue for this circular threshold is around 0.83, so acceptablecircular objects will be selected and simultaneously theperformance of this algorithm will be sufficient enough. Inthis way, all regions with a circular index higher than 0.83will be shown as heating wires as a very important factor forfinal evaluation. The simulation result of this step, which isabout boundaries and circular object indication, are shownin Figure 8. As shown in this figure, circular index for allobjects is determined.
Based on presented information in Figure 8, there arevarious objects with circular index higher than 0.83 and ROIwill be selected around these objects.
Extracted ROI is shown in Figure 9. This region will beincluded by the exact places of the EF joints in polyethylenepipelines.
4.3. Data Extraction by Wavelets Transform. Wavelet trans-form exploits both the spatial and frequency correlationof data by dilations (or contractions) and translations ofmother wavelet on the input data. It supports the multi res-olution analysis of data that can be applied to different scalesaccording to the details required, which allows progressivetransmission and zooming of the image without the need ofextra storage [17, 18].
The implementation of wavelet compression scheme isvery similar to the subband coding scheme: the signal isdecomposed using filter banks (Figure 10). The output of thefilter banks is downsampled, quantized, and encoded. Thedecoder decodes the coded representation, up-samples andrecomposes the signal [17].
Wavelet transform divides the information of an imageinto approximation and detail subsignals. The approxima-tion of sub-signal shows the general trend of pixel values, andother three detail sub-signals show the vertical, horizontaland diagonal details or changes in the images.
In this paper 3D wavelet is used, and by utilizing the ROI(from previous section), all information about images will beextracted by wavelet transform. The results of using waveletare shown in Table 1, where Fi is a matrix for saving theextracted data and all data will indicate the entire extractedfeatures from the images. In fact, in this simulation 5-levelHAAR wavelet is used and it is clear that Fi matrix includesfrom 5 rows and each row will be about each level extractionfor wavelet. Also in this matrix, it is clear that we have 6columns and each column will include features such as: themean and standard deviation of data for horizontal, vertical,and diagonal details, respectively.
4.4. Final Evaluation by ANN. Using the extracted featurevector representations from previous section, neural classi-fier is trained and tested to recognize and classify the scenes.
Neural networks are based on models of biologicalneurons and form a parallel information processing array
Journal of Control Science and Engineering 5
(a) (b)
Figure 4: Image resizing.
(a) (b)
Figure 5: Noise reduction for kind of sample image. (a) before (b) after.
00000
00 0
0
000
111
1111
1 11
111
Figure 6: Structuring element in morphological application.
based on a network of interconnected elements [19, 20]. AMultilayer perceptron (MLP) networks is used in this paper.This type of network is trained using a process of supervisedlearning in which the network is presented with a seriesof matched input and output patterns and the connectionstrengths or weights of the connections are automaticallyadjusted to decrease the difference between the actual anddesired outputs [19]. The structure of this kind of network isshown is Figure 11.
The introduced ANN was trained by the main features ofimages for different joints. So part of the data will be usedfor training phase, and in the next step trained network willbe tested. Figure 12 depicts the converging training graphof neural classifier for Haar wavelet features, respectively.
This network is trained after 4000 epoch and the error isacceptable. After that, a testing set by other images will beused to check the classification performance and its accuracy.
In fact, the trained network is used for testing, and in thisway 10 practical images are applied for this evaluation. Basedon the simulation results, 9 images are recognized correctly.It is assumed that the image with correct joint is 1 and eachimage with incorrect joint is −1.
Based on used testing images we have 7 images with cor-rect joint and 3 incorrect joint.
In this section, based on the previous information, thenetwork is trained. Next, the results of testing network arepresented in Figure 12 and Table 2. Note that the desired
6 Journal of Control Science and Engineering
Sample image 2
Sample image 1(a) (b)
(a) (b)
Figure 7: The effect of morphology on two sample images. (a) Before morphology application, (b) after morphology.
ROI
Metrics closer to 1 indicate that the object is approximately round
Figure 8: Circularity of objects.
Figure 9: Extracted ROI with improved quality.
Journal of Control Science and Engineering 7
LL
LH
HL
HH
LL
LH
HL
HH
LL
LH
HL
HH
LL
LH
HL
HH
LL
LH
HL
HH
LL
LH
HL
HH
LL
LH
HL
HH
LL
LH
HL
HH
LL
LH
HL
HH
LL
LH
HL
HH
LL
LH
HL
HH
LL
LH
HL
HH
LL
LH
HL
HH
LL
LH
HL
HH
LL
LH
HL
HH
↓ 2
↓ 2
↓ 2
↓ 2
↓ 2
↓ 2
↓ 2↓ 2
↓ 2
↓ 2
↓ 2↓ 2
↓ 2
↓ 2
↓ 2
↓ 2
↓ 2
↓ 2
↓ 2↓ 2
↓ 2
↓ 2
↓ 2↓ 2
↓ 2
↓ 2
↓ 2↓ 2
↓ 2
↓ 2
↓ 2↓ 2
↓ 2
↓ 2
↓ 2
↓ 2
↓ 2
↓ 2
↓ 2
↓ 2
↓ 2
↓ 2
↓ 2
↓ 2
↓ 2
↓ 2
↓ 2
↓ 2
↓ 2
↓ 2
↓ 2
↓ 2
↓ 2
↓ 2
↓ 2↓ 2
↓ 2
↓ 2
↓ 2
↓ 2
↓ 2
↓ 2
aL
aL
aL
aL
aL
aL
aL
aL
aL
aL
aL
aL
aL
aL
aL
aL
aL
aL
aL
aL
aL
aL
aL
aL
aL
aL
aL
aL
aL
aL
aLaH
aH
aH
aH
aH
aH
aH
aH
aH
aH
aH
aH
aH
aH
aH
aH
aH
aH
aH
aH
aH
aH
aH
aH
aH
aH
aH
aH
aH
aH
aH
Figure 10: Example of five-level in wavelet transform.
...
...
.........
Hidden layer
Output layer
True
Fault
12345
Input layer
i
Figure 11: A typical multilayer perceptron network.
Classification of correct and incorrect jointQ Actual ANN
Value Type Value Type
1 −1 Disjoint −0.8467 Disjoint OK
2 −1 Disjoint −0.9962 Disjoint OK
3 −1 Disjoint −0.9756 Disjoint OK
4 1 Joint 0.9336 Joint OK
5 1 Joint 0.8449 Joint OK
6 1 Joint 0.9450 Joint OK
7 1 Joint 0.5027 Joint OK
8 1 Joint −0.2186 Disjoint —
9 1 Joint 0.1566 Joint OK
10 1 Joint 0.9777 Joint OK
2
1.8
1.6
1.4
1.2
1
0.8
0.6
0.4
0.2
00 500 1000 1500 2000 2500 3000 3500 4000
Iterations
Err
or
Performance of Net2
Figure 12: Training of network.
outputs for training of networks are −1 and 1. In fact, whenthe estimated outputs of ANN are positive, the PE joint iscorrect but for the negative values the PE joint is in a weakcondition.
5. Conclusion
Based on practical experiments, a new combined algorithmbased on a digital image processing, wavelet transform, arti-ficial neural networks, and mathematical morphology tech-niques is presented.
This algorithm is applied and tested for improvingthe quality of the raw UT images from NDT evaluations.Obtained results show that the proposed algorithms to
ultrasonic signals are highly reliable and also precise forthe NDT testing and monitoring. It should be noted thatextracted ROI by this method is applied for final evaluationby ANN and wavelet transform.
References
[1] J. Shi, J. Zheng, and W. Guo, “Formation mechanism andmicro-structure of the Eigen-line in electro-fusion joints ofpolyethylene pipes,” in Proceedings of the American Society ofMechanical Engineers, Pressure Vessels and Piping Division, pp.303–313, July 2008.
[2] D. J. Hughes, E. L. Heeley, and C. Curfs, “A non-destructivemethod for the measurement of residual strains in semi-crystalline polymer components,” Materials Letters, vol. 65, no.3, pp. 530–533, 2011.
[3] H. Shin, Y. Jang, J. Kwan, and E. Lee, Nondestructive Testingof Fusion Joints of Polyethylene Piping By Real Time UltrasonicImaging, vol. 10, NDT.net, 2005.
[4] H. Kasban, O. Zahran, H. Arafa, M. El-Kordy, S. M. S. Elaraby,and F. E. Abd El-Samie, “Welding defect detection from radio-graphy images with a cepstral approach,” NDT and E Inter-national, vol. 44, no. 2, pp. 226–231, 2011.
[5] H. A. Mehrabi and J. Bowman, “Electrofusion welding ofcross-linked polyethylene pipes,” Iranian Polymer Journal, vol.6, no. 3, pp. 195–203, 1997.
[6] Z. Song, Q. Wang, X. Du, and Y. Wang, “A high speed digitalultrasonic flaw detector based on PC and USB,” in Proceedingsof the IEEE Instrumentation and Measurement Technology(IMTC ’07), May 2007, paper no. 4258079.
[7] J. Brase, R. McKinney, K. Blaedel, J. Oppenheimer, S. Wong,and J. Simmons, “An automated ultrasonic testing bed,”Materials Evaluation, vol. 42, no. 12, pp. 1619–1625, 1984.
[8] I. L. Iznita, S. A. Vijanth, P. A. Venkatachalam, and S.N. Lee, “Computerized segmentation of sinus images,” in
Journal of Control Science and Engineering 9
Proceedings of the Innovative Technologies in Intelligent Systemsand Industrial Applications (CITISIA ’09), pp. 125–128, July2009.
[9] Z. Li, X. Xu, X. Zhu, and C. Sui, “Computer system of ultra-sonic signal sampling and analyzing,” Nondestructive Testing,vol. 17, no. 6, pp. 154–156, 1996.
[10] M. I. H. Bhuiyan, M. O. Ahmad, and M. N. S. Swamy, “Spa-tially adaptive thresholding in wavelet domain for despecklingof ultrasound images,” IET Image Processing, vol. 3, no. 3, pp.147–162, 2009.
[11] A. M. Martı́nez, “Recognizing imprecisely localized, partiallyoccluded, and expression variant faces from a single sampleper class,” IEEE Transactions on Pattern Analysis and MachineIntelligence, vol. 24, no. 6, pp. 748–763, 2002.
[12] S. Mitra and Y. Liu, “Local facial asymmetry for expressionclassification,” in Proceedings of the IEEE Computer SocietyConference on Computer Vision and Pattern Recognition (CVPR’04), pp. II889–II894, June 2004.
[13] M. Vetterli and J. K. Evic, Wavelets and Subband Coding,Prentice-Hall, Englewood Cliffs, NJ, USA, 1995.
[15] Y. Q. Fu and Y. S. Wang, “Algorithm for edge detection ofgray-scale image based on mathematical morphology,” Journalof Harbin Engineering University, vol. 10, no. 5, pp. 685–687,2005.
[16] J. Xu, “Morphological representation of 2-D binary shapesusing rectangular components,” Pattern Recognition, vol. 34,no. 2, pp. 277–286, 2001.
[17] J. S. Walker, A Primer on Wavelets and Their Scientific Appli-cations Edition, Chapman & Hall/CRC, New York, NY, USA,2nd edition, 2008.
[18] F. Bettayeb, T. Rachedi, and H. Benbartaoui, “An improvedautomated ultrasonic NDE system by wavelet and neuronnetworks,” Ultrasonics, vol. 42, no. 1–9, pp. 853–858, 2004.
[19] S. Haykin, Neural Networks and Learning Machines, PearsonEducation, Upper Saddle River, NJ, USA, 3rd edition, 2009.
[20] T. H. Martin, B. D. Howard, and H. B. Mark, Neural NetworkDesign, PWS, Boston, Mass, USA, 1995.
![2014 JCSE ICT Skills Survey - Overviewfilename]_2.pdf · Other SETAs ... this slippage, ranging from limited numbers of matriculants and graduates in the ... 2014 JCSE ICT Skills](https://static.documents.pub/doc/80x56/5ab8c6107f8b9ad3038d36ed/2014-jcse-ict-skills-survey-overview-filename2pdfother-setas-this-slippage.jpg)
